BCLIP-ADer

Abstract

As an essential subsystem of electrified railway operation and maintenance, intelligent detection of catenary support components still faces several critical challenges: (1) the number of abnormal (negative) samples for components is severely limited; (2) component anomalies are highly diverse and exhibit heterogeneous visual characteristics; and (3) existing models generally show unsatisfactory detection performance when confronted with previously unseen anomaly types. To address these issues, this paper proposes a novel few-shot anomaly detection model for catenary components, termed BCLIP-ADer, built upon a Bayesian prompt contrastive vision-language pretraining framework. Specifically, a Bayesian prompt flow module (PFM) is designed to regularize the text prompt space via the jointly learned image-specific feature distribution (ISFD) and image-agnostic feature distribution (IAFD), thereby mitigating the degradation in detection performance on unseen component anomalies. Monte Carlo sampling over these learned distributions is further employed to generate diverse text prompts, leading to more comprehensive coverage of the prompt space. In addition, a cross-modal feature refinement module (CFRM) is designed to more effectively align dynamic text embeddings with fine-grained image features, thus enhancing anomaly detection at the component level. Finally, extensive experiments conducted on a UAV-based catenary dataset (CSCUD) demonstrate the effectiveness and superiority of the proposed approach. Specifically, the proposed method achieves I-AUROC/I-AP/I-F1_max scores of 94.2/93.2/93.1 under few-shot conditions.